Anthropic’s Fable 5.1 is a bit cheaper, a bit smarter, and refuses a lot less

Anthropic rolled out new versions of its flagship Fable and Mythos models on Tuesday, and the pitch is simple: more capability for less money. The company says the updates should perform better while cutting costs, mainly by slashing cache read pricing. The base token prices stay at $10 and $50 per million input and output tokens, but cache reads are now $0.25 per million, down 75%.

Fable 5.1 and Mythos 5.1 are the same model under the hood, but they are not being pushed out in the same way. Fable 5.1 is generally available. Mythos 5.1 stays inside Anthropic’s trusted access program because it includes safeguards the company says are aimed at cybersecurity and life sciences work.

The numbers Anthropic shared suggest a real step forward, though not a dramatic one across the board. The company says Fable 5.1 beats Fable 5, Opus 5, and OpenAI’s GPT-5.6 Sol in its benchmark set, and that many gains come from using lower reasoning settings while still matching older results. One exception stands out: on Terminal-Bench-Science 0.1, which measures agent use in scientific research, Fable 5.1 scored 52.6% versus 24.7% for Fable 5.

That performance bump comes with a softer touch on safety filters, at least for ordinary users. Anthropic says the model’s safeguards are now more precise, so they should block fewer benign requests. In Claude Code, cyber safeguards intervene about 60% less often per session than before, and the model can now identify vulnerabilities in code. Penetration testing, exploit generation, and binary scanning are still blocked and routed to Opus models. For biology, Anthropic says safeguards now intervene 85% less often on benign questions.

There are also changes aimed at enterprise customers, regulators, and anyone suspicious of model reuse. Anthropic is opening a watermark detection API in private preview for eligible organizations under EU law, including regulators, law enforcement, media, fact-checkers, researchers, educational organizations, and EU civil society groups. On retention, the company is moving toward Enterprise Frontier Safeguards, which it says will give customers zero data retention while still allowing monitoring for misuse. The data stays in customer-controlled cloud storage, not Anthropic’s, and the rollout starts later this fall. Fable 5.1 also adds stronger anti-distillation measures by blocking manual edits to Claude’s prior context in certain multi-turn chats, a move that will also affect some agent harnesses.